dm_nfnet_f0 export to onnx, help

[Splendon]

Hi,

I trying to export dm_nfnet_f0 to onnx using gen-efficientnet-pytorch. But raise onnx operator is not be supported.
RuntimeError: Exporting the operator std_mean to ONNX opset version 11 is not supported. Please open a bug to request ONNX export support for the missing operator.
whet can I do to solve this problem, or is there another ways to convert pth.tar(dm_nfnet_f0) to onnx?
Thanks.

Traceback (most recent call last):
File "onnx_export.py", line 83, in
main()
File "onnx_export.py", line 60, in main
torch_out = torch.onnx._export(
File "/home/chenwh/anaconda3/lib/python3.8/site-packages/torch/onnx/init.py", line 28, in _export
result = utils._export(*args, **kwargs)
File "/home/chenwh/anaconda3/lib/python3.8/site-packages/torch/onnx/utils.py", line 632, in _export
_model_to_graph(model, args, verbose, input_names,
File "/home/chenwh/anaconda3/lib/python3.8/site-packages/torch/onnx/utils.py", line 417, in _model_to_graph
graph = _optimize_graph(graph, operator_export_type,
File "/home/chenwh/anaconda3/lib/python3.8/site-packages/torch/onnx/utils.py", line 203, in _optimize_graph
graph = torch._C._jit_pass_onnx(graph, operator_export_type)
File "/home/chenwh/anaconda3/lib/python3.8/site-packages/torch/onnx/init.py", line 263, in _run_symbolic_function
return utils._run_symbolic_function(*args, **kwargs)
File "/home/chenwh/anaconda3/lib/python3.8/site-packages/torch/onnx/utils.py", line 930, in _run_symbolic_function
symbolic_fn = _find_symbolic_in_registry(domain, op_name, opset_version, operator_export_type)
File "/home/chenwh/anaconda3/lib/python3.8/site-packages/torch/onnx/utils.py", line 888, in _find_symbolic_in_registry
return sym_registry.get_registered_op(op_name, domain, opset_version)
File "/home/chenwh/anaconda3/lib/python3.8/site-packages/torch/onnx/symbolic_registry.py", line 111, in get_registered_op
raise RuntimeError(msg)
RuntimeError: Exporting the operator std_mean to ONNX opset version 11 is not supported. Please open a bug to request ONNX export support for the missing operator.

[rwightman]

@Splendon if you google for it you'll find that op was only added to PyTorch recently and not in the last release. You have a few options

• use a release candidate of PyTorch 1.8 or the nightly
• changing the boolean use_layernorm to default to True might work, it hijacks the layer_norm kernel to do the mean/std and might be supported
• you can re-write the layer to use separate mean + std ops

[ooodragon94]

@rwightman another question on "dm_nfnet_f0"
using pip installed timm
seems like there is no model defined on timm version 0.3.4

is there something I am missing or maybe my system is outdated?

thanks

[linhduongtuan]

Many times people ask the same question! Ross does not so often update pip version until he tests carefully ==> you should use the most recent/update version via "pip install git+https://github.com/rwightman/pytorch-image-models.git <git+https://github.com/rwightman/pytorch-image-models.git>” Alternatively, you can download the branch master to your computer and using his train.py. Linh

…

On 23 Feb 2021, at 16:32, ooodragon94 ***@***.***> wrote: @rwightman <https://github.com/rwightman> another question on "dm_nfnet_f0" using pip installed timm seems like there is no model defined on timm version 0.3.4 is there something I am missing or maybe my system is outdated? thanks — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#442 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AFKZ27ES337KT2KGLZ3LGQTTANY3FANCNFSM4X4OEFJQ>.

[linhduongtuan]

@ross, I think you should write some statements/introduction how to install your repo via pip version or/and source code. Linh

…

On 23 Feb 2021, at 16:32, ooodragon94 ***@***.***> wrote: @rwightman <https://github.com/rwightman> another question on "dm_nfnet_f0" using pip installed timm seems like there is no model defined on timm version 0.3.4 is there something I am missing or maybe my system is outdated? thanks — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#442 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AFKZ27ES337KT2KGLZ3LGQTTANY3FANCNFSM4X4OEFJQ>.

[ooodragon94]

@linhduongtuan thank you very much!

[rwightman]

@linhduongtuan @ooodragon94 yes, I should add that git install command in the README. I tend to delay the official pip releases to give the master a bit of time to ensure bugs aren't found (and to allow me to re-run full model validation results.csv generation after adding new models, which takes multiple days now). A bunch of new updates coincided with all the NFNet activity. And my validation run keeps crashing the old test machine so it's taking a while :/

[ooodragon94]

@rwightman always appreciate your hard work. Thx a ton

[Splendon]

@rwightman thank you very much for your suggestion, I'am trying these suggestion to deal with this operator.

[Splendon]

@rwightman I have transformed nfnet from pth to onnx successfully, by using pytorch1.8.